We evaluated the performance of three global evapotranspiration (ET) models at local, regional, and global scales using the multiple sets of leaf area index (LAI) and meteorological data from 1982 to 2017 and investigated the uncertainty in ET simulations from the model structure and forcing data. The three ET models were the Simple Terrestrial Hydrosphere model (SiTH) developed by our team, the Priestley–Taylor Jet Propulsion Laboratory model (PT-JPL), and the MODerate Resolution Imaging Spectroradiometer (MODIS) ET algorithm (MOD16). Comparing the observed with simulated monthly ET by the three models over 43 Fluxnet sites, we found that SiTH overestimated ET for forests with mean slope from 1.25 to 1.67, but it performed better than the other two models over short vegetation. MOD16 and PT-JPL models simulated well for forests but poorly in dryland biomes (slope = 0.25~0.55; R2 = 0.02~0.46). At the catchment scale, all models performed well, except for some tropical and high latitudinal catchments, with NSE values lower than 0 and RMSE and MAE values far beyond their mean values. At the global scale, SiTH highly overestimated ET in tropics, while PT-JPL slightly underestimated ET between 30°N and 60°N and MOD16 underestimated ET between 15°S and 30°S. Generally, the PT-JPL provided the better performance than SiTH and MOD16 models. This study also revealed that the estimated ET by SiTH and especially PT-JPL model were influenced by the uncertainty in meteorological data, and the estimated ET was performed better using MERRA-2 datasets for PT-JPL and using ERA5 datasets for SiTH. While the estimated ET by MOD16 were relatively sensitive to LAI data. In addition, our results suggested that the GLOBMAP and GIMMS datasets were more suitable for long-term ET simulations than the GLASS dataset.